deqing commited on
Commit
1b37551
·
verified ·
1 Parent(s): 7870d88

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +8 -7
README.md CHANGED
@@ -1,20 +1,20 @@
1
  # Qwen3-Inspired Pre-training Dataset
2
 
3
  ## Overview
4
- This is a demo version with sampled data from each source.
5
 
6
  This dataset is a curated mixture of high-quality text data designed for large language model pre-training, inspired by the Qwen3 methodology.
7
 
8
  ## Dataset Statistics
9
 
10
- **Total Size:** 6.78 billion tokens
11
 
12
  ### Data Sources
13
 
14
- - **dclm_baseline**: 3.71B tokens (54.78%) - 2,981,871 documents
15
- - **mini_pile**: 1.43B tokens (21.09%) - 998,100 documents
16
- - **the_stack**: 0.82B tokens (12.09%) - 148,195 documents
17
- - **common_corpus**: 0.82B tokens (12.04%) - 194,877 documents
 
18
 
19
 
20
  ## Data Processing Pipeline
@@ -48,9 +48,10 @@ dataset = load_dataset("bluelightai-dev/qwen_clt_pretrain_data")
48
 
49
  The dataset combines data from the following sources:
50
  - **DCLM Baseline**: High-quality web text from DataComp-LM
51
- - **Common Corpus**: Curated common crawl data
52
  - **The Stack**: Deduplicated source code
53
  - **Mini Pile**: Academic and reference texts
 
54
 
55
  ## License
56
 
 
1
  # Qwen3-Inspired Pre-training Dataset
2
 
3
  ## Overview
 
4
 
5
  This dataset is a curated mixture of high-quality text data designed for large language model pre-training, inspired by the Qwen3 methodology.
6
 
7
  ## Dataset Statistics
8
 
9
+ **Total Size:** 10.19 billion tokens
10
 
11
  ### Data Sources
12
 
13
+ - **dclm_baseline**: 6.21B tokens (60.92%) - 4,973,695 documents
14
+ - **mini_pile**: 1.43B tokens (14.04%) - 999,249 documents
15
+ - **common_corpus**: 1.01B tokens (9.87%) - 246,160 documents
16
+ - **the_stack**: 0.96B tokens (9.4%) - 248,650 documents
17
+ - **math_pile**: 0.59B tokens (5.77%) - 66,729 documents
18
 
19
 
20
  ## Data Processing Pipeline
 
48
 
49
  The dataset combines data from the following sources:
50
  - **DCLM Baseline**: High-quality web text from DataComp-LM
51
+ - **Common Corpus**: Multilingual web text corpus
52
  - **The Stack**: Deduplicated source code
53
  - **Mini Pile**: Academic and reference texts
54
+ - **Math Pile**: Mathematical content and reasoning datasets
55
 
56
  ## License
57